AI/ robotics · humanoid · motion-planning · ai

Humanoid Robot Learns to Adapt Mid-Motion, Not Just Follow a Script

A new planning framework called ReactiveBFM lets humanoid robots recover from physical errors in real time, a step beyond rigid pre-programmed motion models.

Humanoid robots just got a little better at not falling over when the world stops cooperating.

Researchers introduced ReactiveBFM, a closed-loop planning and control framework designed to fix a core weakness in today's humanoid motion systems. Current behavior models execute pre-defined reference motions — fine when conditions match expectations, brittle when they don't. Naively plugging a generative motion planner on top doesn't solve the problem; tiny tracking errors compound until the robot fails. ReactiveBFM addresses this with a training technique called scheduled prefix sampling, which forces the planner to learn recovery behaviors from imperfect physical states rather than clean ground-truth data. An asynchronous replanning mechanism handles the timing gap between slow high-level planning and fast low-level motor control, while trajectory chunking smooths execution to avoid jitter.

The 93.1% success rate under severe perturbations in simulation — a 28.6-point improvement over open-loop baselines — is the number worth watching. More telling is the zero-shot moving-target demo on a Unitree G1 humanoid: the robot tracks a target it was never explicitly trained on, adjusting its whole body in real time. That kind of generalization is what separates a research curiosity from something that might eventually work in a warehouse or a home.

Humanoid robotics is crowded with ambition right now, but most demos still show robots succeeding in controlled conditions. A framework that actively trains on failure states is a more honest engineering approach — though sim-to-real gaps mean the 93.1% figure should be read as a ceiling, not a floor, until field results follow.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →